Kuaishou says Kling 4.0 goes live in October. The company announced the model on September 28 and opened a limited early build of Kling 4.0 Flash the same day.
Two numbers lead the release. Clips get twice as long. The top output format picks up high dynamic range.
The list matters less than which parts of it are switched on. Kuaishou’s own wording dates some of them to a later, unnamed release. Here is the announcement read closely.
What changes from 3.0
Kling 3.0 is what the site sells right now. Its product page still promises “up to 15 seconds” of video from one generation. That is the baseline for the table below.
| Spec | Kling 3.0 | Kling 4.0 |
|---|---|---|
| Longest single generation | 15 seconds | 30 seconds |
| Top resolution | 1080p | 4K and 1080p |
| 10-bit HDR output | Not offered | Announced, marked coming soon |
| Widest aspect ratio | 16:9 | 21:9 added |
| Prompt length | Not stated | Up to 8,000 tokens |
| Reference inputs | Not stated | 10 images, 5 videos, 7 subjects |
| Keyframes | Not stated | Up to 10 |
| Longest extended sequence | Not stated | Up to 2 minutes, marked coming soon |
Only two rows carry an explicit delay. The rest are described as part of the model, with no note about a staged rollout. That is worth remembering when October arrives.
Resolution deserves a reality check. Most platforms re-encode whatever you upload. Almost nobody watches at native 4K. The spec pays off if you deliver to a client or cut for a large screen. For a social clip it buys you crop room, not sharpness.
The 4K line was already on the site
Kling’s home page runs a banner this week that reads “Cinema-Grade Native 4K, Now on Kling AI”, sitting above “World’s First Native 4K Video Model”.
Those lines are older than the announcement. The same two sit on the Kling 3.0 product page, which is still live and still sells the previous model. So the 4K headline is not new, and it is not specific to 4.0.
What 4.0 adds is 10-bit HDR at 4K and 1080p. Kuaishou tags that feature as coming soon in the same paragraph that announces it.
The difference between the two claims is easy to state. 4K sets how many pixels sit in the frame. HDR sets how much of the bright and dark range survives the edit. One of them ships in October. The other is a promise.
A second detail points the same way. Kling’s own structured data, the block that search engines read, still describes the product as the “Kling 3.0 model series” with “15-second multi-shot control”. The press release moved faster than the site it points at.
Thirty seconds, and a number that does not add up
Thirty seconds in one pass is the change most people will feel. It doubles the ceiling, and it moves Kling from fragment length into something that fits a full ad beat.
The release also adds up to 10 keyframes. You can pin specific moments and let the model fill the space between them. Both changes point the same way. Fewer separate generations to stitch afterward.
Two smaller items sit in the same section. The first is 21:9. That is the shape most film and trailer work uses, and until now you had to crop 16:9 output to fake it. Cropping throws away pixels. Generating in the wider frame does not.
The second is prompt length, now up to 8,000 tokens. That is roughly six thousand words. Nobody types that. What it buys is room for a long structured brief, with camera notes and character descriptions pasted in whole.
Then comes the reference paragraph, and it has a problem. Kuaishou describes “up to 15 multimodal references” and breaks that down as 10 images, 5 videos and 7 subjects. Ten plus five plus seven is 22. Chinese outlets that carried the release repeat both numbers without reconciling them.
The reading that fits is that 15 counts images and videos, with the 7 subjects held as a separate cap on recurring characters or objects. But Kuaishou has not put that in writing. If you plan a project around the reference limit, budget for the smaller number.
Audio, accents and the rest
Kling 4.0 keeps the native audio from 3.0 and adds stereo. The release describes two channels rather than one, and tighter lip-sync against dialogue.
The accent list is the part aimed at a US audience. Kuaishou names American, British and Indian English alongside Mandarin, Cantonese, Japanese, Korean, Spanish, Portuguese, German, French and Hindi. It also claims stable text and logo rendering inside moving footage. That is the feature that decides whether a product ad is usable or needs a reshoot.
Two more items sit on the page rather than in the headline. The generator takes video as an input for editing, so you can add, change or remove a subject or a background. And the new model can read the shot structure of a reference clip, then build a commercial in that shape.
How you can try it before October
Flash is the lighter, cheaper model, and it went to a small test group on September 28. The Chinese announcement says the first group is made up of “black gold” annual pass holders.
That name does not appear on the English site. kling.ai sells five plans called Basic, Standard, Pro, Premier and Ultra. Nothing in the release maps the Chinese membership onto those five. So if you hold a paid plan and want Flash early, the announcement does not tell you where you stand, and there is no waitlist to join.
The home page currently runs a discount banner on annual plans, with a countdown and a spot counter. Treat that as an ordinary promotion. Kuaishou has not connected it to 4.0 access anywhere.
What one 4.0 job costs
Kling bills in credits, so a longer clip carries a price.
The compose box on the Kling 4.0 page shows 70 credits against its default settings of 480p at 16:9. The label does not say whether that figure is the cost of the job or the balance on the account.
The 4K headline and the 480p default sit on the same page.
The membership page prices things on a different scale. It still describes the Standard plan as roughly 33 short clips a month, which works out near 20 credits each. Both numbers live on the same site, and neither page says which model it prices.
The gap is worth knowing before you top up for a project.
The launch this was aimed at
The timing is not accidental. Shanghai Securities News, reporting the release, frames it as a response to Seedance 2.5 from ByteDance, the company that also makes CapCut. That model launched in July and drew strong reviews. Alibaba and MiniMax are named as pressure from the cheaper end of the market.
Kling has money for that fight. Kuaishou announced a funding round of about $3 billion for the Kling unit in July, at a valuation near $18 billion, with a five-year listing commitment attached. The same report notes that early reaction in China was muted, with commenters asking about price and capability rather than praising the demos.
None of that changes the specs. It does explain why the release leads with length and format instead of a cheaper tier.
What is still missing
A month is as specific as Kuaishou got. There is no day, no price sheet for the new model, and no word on when HDR clears or extended sequences ship.
When October lands, the questions worth asking are the ones the announcement skipped. Does the free tier keep its watermark on 4.0 output. Do the credits you already own spend on 4.0 jobs. And does the new model cost more per second than 3.0 does today.
The CapCut page runs this same check on another app. And the DeepSeek page repeats it for a text model. The ten-app availability table puts Kling next to the rest, and all of it sits under the China Tech in America index.
Thirty seconds is the claim to test. If a single pass really holds that long without falling apart, it removes a step from every project that used to need two clips stitched together. That is more useful than the resolution line, and it is the one worth checking in October.
